Guiding a Harsh-Environments Robust Detector via RAW Data Characteristic Mining

Authors: Hongyang Chen, Hung-Shuo Tai, Kaisheng Ma

AAAI 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Specifically, our experiments indicate that PRD (using FCOS) outperforms RGB detection by 13.9m AP on LOD-Snow without generating restored images.
Researcher Affiliation Collaboration 1Xi an Jiaotong University, Xi an, China 2Kargo Bot.ai, Beijing, China 3Tsinghua University, Beijing, China chenhy@stu.xjtu.edu.cn, hungshuotai@didiglobal.com, kaisheng@mail.tsinghua.edu.cn
Pseudocode No The paper describes methods and includes mathematical formulas but does not contain a clearly labeled 'Pseudocode' or 'Algorithm' block.
Open Source Code Yes The code is available at https://github.com/Dreamer CCC/Raw Mining.
Open Datasets Yes To evaluate the real-world performance of low-light detection, we utilize the LOD (Hong et al. 2021) dataset, which consists of 2230 image pairs that are randomly split into a training set of 1830 pairs and a test set of 400 pairs. The RAW-NOD (Morawski et al. 2022) dataset contains 7K raw images captured in outdoor low-light conditions. PASCALRAW (Omid-Zohoor, Ta, and Murmann 2014) contains 4,259 annotated RAW images, with three annotated object classes (car, person, and bicycle), and is modeled after the PASCAL VOC database.
Dataset Splits No The paper states 'randomly split into a training set of 1830 pairs and a test set of 400 pairs' for the LOD dataset, but does not explicitly mention a validation split with specific numbers or percentages.
Hardware Specification Yes running on 8 RTX NVIDIA 2080Ti GPUs (12GB).
Software Dependencies No The paper mentions 'Open MMLab Detection Toolbox (Chen et al. 2019) and PyTorch' but does not specify their version numbers.
Experiment Setup Yes We follow the official default settings of detectors, e.g. for Center Net, use Random Center Crop Pad and Random Flip as data augmentation.